Research On Cardiovascular Disease Prediction Orienting Complex Factors

Cardiovascular disease(CVD)acts as the number one killer of urban and rural residents,it is strongly demand for primary screening of CVD in the grassroots community.Formulating targeted intervention measures on the basis of CVD risk prediction can effectively reduce the incidence of the disease.The core technical link in CVD prediction is establishing prediction model with high performance.The development of economic and social has pushed people's life style and environment changed rapidly,which makes the CVD risk factors become more complex.Hence it is of great practical value and practical significance to establish a new CVD prediction model orienting complex factors.Conventional CVD prediction is using logistic regression,cox regression and other mathematical forecasting methods to establish formulaic prediction models.Constrained by linear fitting capacity of formula,all factors included in these models are continuous and binary classification variables,and the polytomous complex factors expressed by nonlinear data can't be applied in them.But the complex factors may contain important potential information,which can contribute to the accurate prediction of CVD.Neural networks have strong learning and adaptive ability in nonlinear data processing,which provides a way to solve this problem.This thesis aims to achieve the CVD prediction orienting complex factors.Firstly,we analyze the specific reasons for why the regression model can't be geared to CVD prediction when orienting complex factors,and then the complex factors are linearized by dummy variable to solve the problem.Secondly,the nonlinear complex factors are mapped to high dimensional space to do linear regression fitting by shallow neural network,which realizes the CVD prediction orienting complex factors and improves the AUC(the area under ROC curve)value of the model.Finally,a CVD prediction method is designed based on deep learning.The initialization strategy of shallow neural network parameters are improved to reduce the variance of predicted results by using unsupervised learning.The main research works and results as follows:(1)From the perspective of model mechanism,we analyze the reason of why regression models can't be applied to complex factors.Then the method of setting dummy variables to complex factors are adopted to solve this problem.Taking logistic regression model as example to do experiment.Under the condition of including complex factors,the AUC value of the improved model is increased from 0.7634,0.6700 of logistic up to 0.8784,0.7999 respectively,which is accord with 0.78~0.86 of traditional regression models.The results indicate that regression models can't be directly used to incorporate into complex factors.(2)According to the data characteristics of CVD complex factors,we establish a CVD prediction model based on shallow neural network,and improve prediction accuracy of the model by improving the network's learning ability.Experimental results show that the average AUC values of the model are increased to 0.9024,0.8423 respectively.(3)The expressive ability of CVD data features are improved stepwise by deep learning.Neural network is initialized by the best parameters to solve the problem of predictive instability caused by random initialization of parameters.By this way,we can reduce the variance of the predicted results and improve the stability of the model.Results showed that variance of the prediction results decreased from 12.665,9.051 of shallow neural network to 5.723,4.642 respectively,AUC values are increased to 0.9198,0.8959 respectively.